CopilotKit open-sources OpenDots, an agent workspace with dedicated computers

OpenDots is an open-source agent workspace template from CopilotKit that lets you run persistent AI coworkers, each equipped with its own sandboxed computer.

It is built as an open, self-hosted alternative to proprietary assistant products. You can hook it up to any OpenAI-compatible model provider or run your own local models if your hardware allows.

What it can do

Instead of a generic chatbot, OpenDots organizes work around specialized agents. A researcher can dig through pages while a writer drafts memos, seamlessly interacting with you across web chat, WebRTC voice calls, or Slack threads.

Each Dot gets its own isolated computer powered by OpenBot container services. It maintains its own browser profile, workspace files, and terminal session across restarts. A dedicated computer panel lets you watch the live browser in action, inspect terminal output, or take over control manually at any time.

Work is organized into document libraries called Spaces. Documents feature rich text editing, slash commands, and autosave. When an agent drafts a page, a human-in-the-loop review card pauses execution until you click Approve & save.

External integrations run through standard MCP servers and Composio. Read-only actions proceed automatically, while any write request triggers an in-chat approval prompt so the model never modifies your real accounts without permission.

How to run it

The project is fully available on GitHub as a template. Running it locally requires Node.js 24 and npm.

Clone the repository, install dependencies with npm ci, and copy the example environment file. The interface runs on React SDK and AG-UI, with TanStack AI handling streaming and tool execution.

To persist conversations and agent threads, you need CopilotKit Intelligence, which can be run locally using Docker or connected to their cloud service. Once you add your model API keys, the workspace is ready to go.

Why it matters

Proprietary assistant services ask you to trust a closed cloud with your browser sessions, emails, and sensitive workplace documents. They also rarely give you granular controls over what the agent can touch.

OpenDots takes a smarter architectural approach by sandboxing each agent into its own container and enforcing human-in-the-loop review on all state-changing actions.

We think this is a compelling starting point for developers and engineering teams who want the convenience of autonomous assistants without surrendering control over their tools and infrastructure.